PPO深度指南

关键词速览

核心概念	信任域	剪裁目标	KL散度	自适应惩罚
自然梯度	价值裁剪	广义优势估计	策略更新	目标函数

核心关键词表

术语英文符号/技术说明
近端策略优化 PPO PPO 稳定策略更新的算法
信任域策略优化 TRPO TRPO 基于KL散度约束的策略优化
剪裁目标函数 Clipped Objective $min (r_{t}, clip)$ 防止策略过大更新
替代优势函数 Surrogate Advantage $r_{t} (θ) \cdot \hat{A}_{t}$ 策略比值的加权
KL散度 KL Divergence $D_{KL}(\pi_{\theta_{old}}
自适应KL惩罚 Adaptive KL Penalty $β$ 自动调整的KL惩罚系数
价值函数裁剪 Value Clipping $clip (V, V_{o l d} - ϵ, V_{o l d} + ϵ)$ 防止价值函数剧变
广义优势估计 GAE $\hat{A}_{t}^{G A E (λ)}$ 偏差-方差平衡的优势估计
目标函数 Objective $L^{C L I P + V F + S}$ PPO综合目标
策略比值 Probability Ratio $r_t(\theta) = \frac{\pi_\theta(a_t s_t)}{\pi_{\theta_{old}}(a_t

术语	英文	符号/技术	说明
近端策略优化	PPO	PPO	稳定策略更新的算法
信任域策略优化	TRPO	TRPO	基于KL散度约束的策略优化
剪裁目标函数	Clipped Objective	$min (r_{t}, clip)$	防止策略过大更新
替代优势函数	Surrogate Advantage	$r_{t} (θ) \cdot \hat{A}_{t}$	策略比值的加权
KL散度	KL Divergence	$D_{KL}(\pi_{\theta_{old}}
自适应KL惩罚	Adaptive KL Penalty	$β$	自动调整的KL惩罚系数
价值函数裁剪	Value Clipping	$clip (V, V_{o l d} - ϵ, V_{o l d} + ϵ)$	防止价值函数剧变
广义优势估计	GAE	$\hat{A}_{t}^{G A E (λ)}$	偏差-方差平衡的优势估计
目标函数	Objective	$L^{C L I P + V F + S}$	PPO综合目标
策略比值	Probability Ratio	$r_t(\theta) = \frac{\pi_\theta(a_t	s_t)}{\pi_{\theta_{old}}(a_t

一、从TRPO到PPO

1.1 TRPO原理

信任域策略优化（Trust Region Policy Optimization, TRPO）由Schulman等人于2015年提出，是一种保证策略单调改进的策略优化方法。TRPO的核心约束是限制相邻策略之间的KL散度：

θ max E_{t} [\frac{π _{θ} ( a _{t} ∣ s _{t} )}{π _{θ_{o l d}} ( a _{t} ∣ s _{t} )} \hat{A}_{t}]

s.t. E_{t} [D_{K L} (π_{θ_{o l d}} (\cdot ∣ s_{t}) ∣∣ π_{θ} (\cdot ∣ s_{t}))] \leq δ

TRPO的核心思想

“信任域”隐喻：想象你在黑暗中山徒步，只能信任手电筒照亮的区域。TRPO同样限制策略更新步长在”可信任”的范围内，确保每次更新都是安全的改进。

1.2 TRPO的计算挑战

TRPO需要求解约束优化问题，通常使用共轭梯度法近似求解自然梯度。这种方法存在以下问题：

计算复杂度高：需要二阶优化，计算量大
内存开销大：需要存储Fisher信息矩阵或使用K-FAC近似
与噪声抑制不兼容：无法轻易与GAE等噪声技术结合
实现困难：自然梯度实现复杂

1.3 PPO的诞生

PPO（Proximal Policy Optimization）由Schulman等人于2017年提出，通过一阶优化和剪裁机制简化了TRPO，同时保持甚至超越了TRPO的性能。PPO的核心洞察是：不需要精确的约束，只需要防止策略更新过大。

二、PPO算法详解

2.1 剪裁替代目标函数

PPO使用剪裁机制限制策略更新的幅度：

L^{C L I P} (θ) = E_{t} [min (r_{t} (θ) \cdot \hat{A}_{t}, clip (r_{t} (θ), 1 - ϵ, 1 + ϵ) \cdot \hat{A}_{t})]

其中策略比值：

r_{t} (θ) = \frac{π _{θ} ( a _{t} ∣ s _{t} )}{π _{θ_{o l d}} ( a _{t} ∣ s _{t} )}

2.2 剪裁机制的几何解释

def compute_ppo_loss(log_probs_old, log_probs_new, advantages, epsilon=0.2):
    """
    PPO clip loss computation.
    
    epsilon: 超参数，通常为0.1或0.2
    """
    # 策略比值
    ratio = torch.exp(log_probs_new - log_probs_old)
    
    # 未剪裁的替代损失
    surr1 = ratio * advantages
    
    # 剪裁后的替代损失
    surr2 = torch.clamp(ratio, 1 - epsilon, 1 + epsilon) * advantages
    
    # 取最小值（当A>0时剪裁下界，当A<0时剪裁上界）
    policy_loss = -torch.min(surr1, surr2).mean()
    
    return policy_loss

剪裁机制详解

考虑两种情况：

优势函数 $A > 0$ （动作好于平均）：

如果 $r_{t} > 1 + ϵ$ ：比值被剪裁到 $1 + ϵ$ ，阻止过度增加该动作概率

$min$ 操作确保我们不会过度乐观地估计收益

优势函数 $A < 0$ （动作差于平均）：

如果 $r_{t} < 1 - ϵ$ ：比值被剪裁到 $1 - ϵ$ ，阻止过度降低该动作概率

这提供了一种”保护”，即使策略很差也能稳定更新

2.3 完整PPO实现

import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import gym
 
class PPOPolicy(nn.Module):
    """
    PPO policy network with Gaussian policy for continuous control.
    """
    def __init__(self, obs_dim, act_dim, hidden_dim=64):
        super().__init__()
        
        # Actor (policy)
        self.actor = nn.Sequential(
            nn.Linear(obs_dim, hidden_dim),
            nn.Tanh(),
            nn.Linear(hidden_dim, hidden_dim),
            nn.Tanh(),
            nn.Linear(hidden_dim, hidden_dim),
            nn.Tanh()
        )
        self.log_std = nn.Parameter(torch.zeros(act_dim))
        
        # Critic (value function)
        self.critic = nn.Sequential(
            nn.Linear(obs_dim, hidden_dim),
            nn.Tanh(),
            nn.Linear(hidden_dim, hidden_dim),
            nn.Tanh(),
            nn.Linear(hidden_dim, hidden_dim),
            nn.Tanh(),
            nn.Linear(hidden_dim, 1)
        )
    
    def forward(self, x):
        return self.actor(x), self.critic(x)
    
    def get_action(self, x, deterministic=False):
        """Sample action from policy."""
        mean = self.actor(x)
        std = self.log_std.exp()
        
        if deterministic:
            action = mean
        else:
            dist = torch.distributions.Normal(mean, std)
            action = dist.sample()
        
        action = torch.tanh(action)
        return action, dist.log_prob(action)
    
    def evaluate_actions(self, x, actions):
        """Compute log probs for actions (for PPO update)."""
        mean = self.actor(x)
        std = self.log_std.exp()
        dist = torch.distributions.Normal(mean, std)
        
        # Actions are already tanh-squashed
        log_probs = dist.log_prob(actions)
        
        # Add tanh correction for log_prob
        action_tanh = torch.tanh(actions)
        log_probs -= torch.log(1 - action_tanh.pow(2) + 1e-6)
        
        return log_probs.sum(-1, keepdim=True)
 
class PPOAgent:
    """
    Proximal Policy Optimization (PPO) agent.
    """
    def __init__(self, obs_dim, act_dim, hidden_dim=64, 
                 lr=3e-4, gamma=0.99, lam=0.95, 
                 clip_epsilon=0.2, value_coef=0.5, entropy_coef=0.0,
                 ppo_epochs=10, batch_size=64,
                 max_grad_norm=0.5):
        
        self.clip_epsilon = clip_epsilon
        self.value_coef = value_coef
        self.entropy_coef = entropy_coef
        self.ppo_epochs = ppo_epochs
        self.batch_size = batch_size
        self.max_grad_norm = max_grad_norm
        self.gamma = gamma
        self.lam = lam
        
        # Network
        self.policy = PPOPolicy(obs_dim, act_dim, hidden_dim)
        self.optimizer = optim.Adam(self.policy.parameters(), lr=lr)
        
        # Memory buffer
        self.buffer = []
    
    def select_action(self, obs, deterministic=False):
        """Select action given observation."""
        obs_tensor = torch.FloatTensor(obs).unsqueeze(0)
        
        with torch.no_grad():
            action, log_prob = self.policy.get_action(obs_tensor, deterministic)
            value = self.policy.critic(obs_tensor)
        
        return action.numpy()[0], log_prob.item(), value.item()
    
    def store_transition(self, obs, action, log_prob, value, reward, done):
        """Store transition in buffer."""
        self.buffer.append({
            'obs': obs,
            'action': action,
            'log_prob': log_prob,
            'value': value,
            'reward': reward,
            'done': done
        })
    
    def compute_gae(self, rewards, values, dones):
        """Compute Generalized Advantage Estimation."""
        advantages = []
        gae = 0
        
        values = [0.0] + values + [0.0]  # Pad values
        
        for t in reversed(range(len(rewards))):
            delta = rewards[t] + self.gamma * values[t + 1] * (1 - dones[t]) - values[t]
            gae = delta + self.gamma * self.lam * (1 - dones[t]) * gae
            advantages.insert(0, gae)
        
        return np.array(advantages)
    
    def compute_returns(self, rewards, dones, gamma=0.99):
        """Compute discounted returns."""
        returns = []
        discounted = 0
        
        for reward, done in zip(reversed(rewards), reversed(dones)):
            discounted = reward + gamma * discounted * (1 - done)
            returns.insert(0, discounted)
        
        return np.array(returns)
    
    def update(self):
        """Perform PPO update from collected buffer."""
        if len(self.buffer) < self.batch_size:
            return {}
        
        # Extract from buffer
        obs = np.array([t['obs'] for t in self.buffer])
        actions = np.array([t['action'] for t in self.buffer])
        old_log_probs = np.array([t['log_prob'] for t in self.buffer])
        values = np.array([t['value'] for t in self.buffer])
        rewards = np.array([t['reward'] for t in self.buffer])
        dones = np.array([t['done'] for t in self.buffer])
        
        # Compute advantages and returns
        advantages = self.compute_gae(rewards, values, dones)
        returns = advantages + np.array(values)
        
        # Normalize advantages
        advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-8)
        
        # Convert to tensors
        obs_tensor = torch.FloatTensor(obs)
        actions_tensor = torch.FloatTensor(actions)
        old_log_probs_tensor = torch.FloatTensor(old_log_probs)
        advantages_tensor = torch.FloatTensor(advantages)
        returns_tensor = torch.FloatTensor(returns)
        
        # PPO update for multiple epochs
        policy_losses = []
        value_losses = []
        entropy_losses = []
        
        dataset_size = len(obs)
        indices = np.arange(dataset_size)
        
        for epoch in range(self.ppo_epochs):
            np.random.shuffle(indices)
            
            for start in range(0, dataset_size, self.batch_size):
                end = start + self.batch_size
                batch_idx = indices[start:end]
                
                batch_obs = obs_tensor[batch_idx]
                batch_actions = actions_tensor[batch_idx]
                batch_old_log_probs = old_log_probs_tensor[batch_idx]
                batch_advantages = advantages_tensor[batch_idx]
                batch_returns = returns_tensor[batch_idx]
                
                # Evaluate actions
                log_probs = self.policy.evaluate_actions(batch_obs, batch_actions)
                
                # PPO policy loss (clipped)
                ratio = torch.exp(log_probs - batch_old_log_probs)
                
                surr1 = ratio * batch_advantages
                surr2 = torch.clamp(ratio, 1 - self.clip_epsilon, 1 + self.clip_epsilon) * batch_advantages
                policy_loss = -torch.min(surr1, surr2).mean()
                
                # Value loss (optional clipping)
                values_pred = self.policy.critic(batch_obs).squeeze()
                value_loss = nn.MSELoss()(values_pred, batch_returns)
                
                # Entropy bonus (if applicable)
                entropy_loss = 0  # Can add entropy regularization here
                
                # Total loss
                loss = policy_loss + self.value_coef * value_loss - self.entropy_coef * entropy_loss
                
                # Gradient step
                self.optimizer.zero_grad()
                loss.backward()
                torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.max_grad_norm)
                self.optimizer.step()
                
                policy_losses.append(policy_loss.item())
                value_losses.append(value_loss.item())
        
        # Clear buffer
        self.buffer.clear()
        
        return {
            'policy_loss': np.mean(policy_losses),
            'value_loss': np.mean(value_losses)
        }
 
def train_ppo(env_name='HalfCheetah-v2', num_steps=1000000, 
              num_env_steps=2048, update_freq=2048):
    """Train PPO agent."""
    env = gym.make(env_name)
    
    obs_dim = env.observation_space.shape[0]
    act_dim = env.action_space.shape[0]
    
    agent = PPOAgent(
        obs_dim=obs_dim,
        act_dim=act_dim,
        hidden_dim=64,
        lr=3e-4,
        gamma=0.99,
        lam=0.95,
        clip_epsilon=0.2,
        value_coef=0.5,
        ppo_epochs=10,
        batch_size=64
    )
    
    obs = env.reset()
    episode_rewards = []
    total_steps = 0
    
    while total_steps < num_steps:
        episode_reward = 0
        episode_steps = 0
        
        for _ in range(num_env_steps):
            action, log_prob, value = agent.select_action(obs)
            next_obs, reward, done, _ = env.step(action)
            
            agent.store_transition(obs, action, log_prob, value, reward, done)
            
            episode_reward += reward
            episode_steps += 1
            total_steps += 1
            
            obs = next_obs
            
            if done:
                episode_rewards.append(episode_reward)
                obs = env.reset()
                episode_reward = 0
        
        # Update
        losses = agent.update()
        
        if total_steps % 10000 == 0:
            avg_reward = np.mean(episode_rewards[-10:]) if episode_rewards else 0
            print(f"Steps: {total_steps}, Avg Reward (last 10): {avg_reward:.2f}")
            print(f"  Policy Loss: {losses.get('policy_loss', 0):.4f}, "
                  f"Value Loss: {losses.get('value_loss', 0):.4f}")

三、自适应KL散度惩罚

3.1 自适应KL机制

除了剪裁目标，PPO还支持自适应KL散度惩罚版本：

L^{K L PEN} (θ) = E_{t} [\frac{π _{θ} ( a _{t} ∣ s _{t} )}{π _{θ_{o l d}} ( a _{t} ∣ s _{t} )} \hat{A}_{t} - β D_{K L} (π_{θ_{o l d}} ∣∣ π_{θ})]

其中惩罚系数 $β$ 自适应调整：

class AdaptiveKLPPO:
    """PPO with adaptive KL penalty."""
    
    def __init__(self, target_kl=0.01, kl_penalty_coef=0.5):
        self.target_kl = target_kl
        self.kl_penalty_coef = kl_penalty_coef
        self.kl_history = []
    
    def update_beta(self, kl_divergence):
        """Adaptively adjust KL penalty coefficient."""
        self.kl_history.append(kl_divergence)
        
        if len(self.kl_history) < 10:
            return self.kl_penalty_coef
        
        recent_kl = np.mean(self.kl_history[-10:])
        
        if recent_kl < self.target_kl * 0.5:
            # KL too small, increase penalty to encourage more change
            self.kl_penalty_coef *= 2
        elif recent_kl > self.target_kl * 2:
            # KL too large, decrease penalty to slow down
            self.kl_penalty_coef /= 2
        
        # Clamp to reasonable range
        self.kl_penalty_coef = np.clip(self.kl_penalty_coef, 1e-4, 1e3)
        
        return self.kl_penalty_coef
    
    def compute_kl_divergence(self, old_policy, new_policy, states):
        """Compute average KL divergence between policies."""
        with torch.no_grad():
            old_log_probs = old_policy.evaluate_actions(states, states)
            new_log_probs = new_policy.evaluate_actions(states, states)
            
            # Approximate KL: D_KL(P||Q) ≈ exp(log(P) - log(Q)) - 1 - (log(P) - log(Q))
            kl = torch.exp(old_log_probs - new_log_probs) - 1 - (old_log_probs - new_log_probs)
        
        return kl.mean().item()

3.2 两种PPO变体对比

特性	PPO-Clip	PPO-Penalty
约束方式	硬剪裁	软惩罚
超参数	$ϵ$	$β, δ_{K L}$
稳定性	更稳定	对 $β$ 敏感
适用场景	通用	需要细粒度控制

四、PPO vs TRPO vs A2C对比

4.1 算法对比表

特性	A2C	TRPO	PPO
优化方式	异步/同步梯度	约束优化	一阶+剪裁
策略约束	无（熵正则化）	KL散度约束	概率比剪裁
计算效率	高	中	高
内存开销	低	中（二阶信息）	低
实现难度	低	高	中
样本效率	中	高	高
收敛稳定性	中	高	高
大规模应用	一般	一般	首选

4.2 为什么PPO成为主流？

PPO的成功因素

实现简单：只需一阶优化，无需复杂二阶计算

超参数鲁棒： $ϵ$ 不需要精细调整

通用性好：在离散和连续动作空间都表现优异

与现有框架兼容：易于与深度学习框架集成

样本效率：PPO 2在样本效率上有显著提升

4.3 PPO2与向量化环境

PPO2利用向量化环境提高样本效率：

import gym
from stable_baselines3.common.vec_env import DummyVecEnv, SubprocVecEnv
 
def make_env(env_id):
    """Create environment factory."""
    def _init():
        env = gym.make(env_id)
        return env
    return _init
 
# 创建向量化环境
def create_vectorized_env(env_id, n_envs=8):
    """
    Create vectorized environments for parallel data collection.
    PPO2 style.
    """
    env = SubprocVecEnv([make_env(env_id) for _ in range(n_envs)])
    return env
 
class PPO2Agent:
    """
    PPO2 with vectorized environments.
    """
    def __init__(self, env_fn, n_steps=128, n_epochs=10, nminibatches=4):
        self.env = env_fn()
        self.n_steps = n_steps  # Steps per update
        self.n_epochs = n_epochs
        self.nminibatches = nminibatches
        
        # Calculate batch/minibatch sizes
        self.n_envs = self.env.num_envs
        self.batch_size = self.n_steps * self.n_envs
        self.minibatch_size = self.batch_size // nminibatches
    
    def collect_rollouts(self):
        """Collect n_steps of data from all environments."""
        observations = []
        actions = []
        rewards = []
        dones = []
        values = []
        log_probs = []
        
        obs = self.env.reset()
        
        for _ in range(self.n_steps):
            obs_tensor = torch.FloatTensor(obs)
            
            with torch.no_grad():
                action, log_prob = self.policy.get_action(obs_tensor)
                value = self.policy.critic(obs_tensor).squeeze()
            
            action_np = action.numpy()
            
            # Step all environments
            next_obs, reward, done, info = self.env.step(action_np)
            
            observations.append(obs)
            actions.append(action_np)
            rewards.append(reward)
            dones.append(done)
            values.append(value.numpy())
            log_probs.append(log_prob.numpy())
            
            obs = next_obs
        
        return (np.array(observations), np.array(actions), 
                np.array(rewards), np.array(dones),
                np.array(values), np.array(log_probs))

五、实战调参经验

5.1 超参数推荐值

PPO超参数调优指南

参数推荐范围调整建议
学习率 3e-4 可使用线性衰减
Clip epsilon 0.1 - 0.3 默认0.2效果良好
GAE lambda 0.9 - 0.99 高值更平滑
PPO epochs 10 - 30 10通常足够
Mini-batch size 32 - 256 64/128常用
价值系数 0.5 - 1.0 0.5是标准
熵系数 0 - 0.01 鼓励探索时加
梯度裁剪 0.5 - 1.0 默认0.5
环境数量 1 - 16 根据CPU调整

参数	推荐范围	调整建议
学习率	3e-4	可使用线性衰减
Clip epsilon	0.1 - 0.3	默认0.2效果良好
GAE lambda	0.9 - 0.99	高值更平滑
PPO epochs	10 - 30	10通常足够
Mini-batch size	32 - 256	64/128常用
价值系数	0.5 - 1.0	0.5是标准
熵系数	0 - 0.01	鼓励探索时加
梯度裁剪	0.5 - 1.0	默认0.5
环境数量	1 - 16	根据CPU调整

5.2 调试技巧

class PPODebugger:
    """Utilities for debugging PPO training."""
    
    @staticmethod
    def check_ratio_distribution(ratios):
        """检查策略比值分布."""
        ratios = np.array(ratios)
        
        print(f"Ratio Stats:")
        print(f"  Mean: {ratios.mean():.3f}")
        print(f"  Std: {ratios.std():.3f}")
        print(f"  Min: {ratios.min():.3f}")
        print(f"  Max: {ratios.max():.3f}")
        print(f"  Clipped (>1+ε): {(ratios > 1.2).mean()*100:.1f}%")
        print(f"  Clipped (<1-ε): {(ratios < 0.8).mean()*100:.1f}%")
        
        return ratios
    
    @staticmethod
    def check_gradient_norm(model):
        """检查梯度范数."""
        total_norm = 0
        for p in model.parameters():
            if p.grad is not None:
                param_norm = p.grad.data.norm(2)
                total_norm += param_norm.item() ** 2
        total_norm = total_norm ** 0.5
        return total_norm
    
    @staticmethod
    def check_value_predictions(values, returns):
        """检查价值预测质量."""
        values = np.array(values)
        returns = np.array(returns)
        
        explained_var = 1 - np.var(returns - values) / np.var(returns)
        
        print(f"Value Function Stats:")
        print(f"  Explained Variance: {explained_var:.3f}")
        print(f"  Value Mean: {values.mean():.3f}")
        print(f"  Returns Mean: {returns.mean():.3f}")
        
        return explained_var

5.3 常见问题与解决

问题	原因	解决方案
策略崩溃	更新过大	减小学习率或 $ϵ$
价值函数不稳定	价值预测噪声大	增加价值系数，减小batch size
探索不足	熵系数太小	增加熵系数
训练发散	梯度爆炸	增加梯度裁剪，检查reward缩放
回报不提升	环境或reward设计问题	简化环境，调试reward

六、数学形式化总结

PPO剪裁目标函数

L^{C L I P} (θ) = E_{t} [min (r_{t} (θ) \hat{A}_{t}, clip (r_{t} (θ), 1 - ϵ, 1 + ϵ) \hat{A}_{t})]

自适应KL惩罚目标

L^{K L PEN} (θ) = E_{t} [r_{t} (θ) \hat{A}_{t} - β D_{K L} (π_{θ_{o l d}} ∣∣ π_{θ})]

PPO完整目标

L^{PPO} (θ) = L^{C L I P} (θ) - c_{1} L^{V F} (θ) + c_{2} S [π_{θ}] (s)

其中 $c_{1}$ 是价值系数， $c_{2}$ 是熵系数， $S$ 是策略熵。

七、相关文档

MDP与Bellman方程 — 理论基础
Q学习 — 值函数方法
DQN — 深度值函数方法
策略梯度方法 — 策略梯度基础
多智能体RL — 多智能体扩展
RL应用场景 — 实际应用案例

参考文献

Schulman, J., Levine, S., Abbeel, P., Jordan, M., & Moritz, P. (2015). Trust region policy optimization. ICML, 1889-1897.
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms. arXiv:1707.06347.
Engstrom, L., et al. (2020). Implementation matters in RL: A case study on PPO. ICLR 2020.
Andrychowicz, M., et al. (2020). What matters in on-policy reinforcement learning? A large-scale empirical study. arXiv:2006.05990.
Raffin, A., et al. (2021). SB3: Stable Baselines3. JMLR.

PPO以其简洁的实现和稳定的性能，成为当前强化学习研究和应用的首选算法。

人工智能知识库

探索

PPO深度指南

PPO深度指南

关键词速览

一、从TRPO到PPO

1.1 TRPO原理

1.2 TRPO的计算挑战

1.3 PPO的诞生

二、PPO算法详解

2.1 剪裁替代目标函数

2.2 剪裁机制的几何解释

2.3 完整PPO实现

三、自适应KL散度惩罚

3.1 自适应KL机制

3.2 两种PPO变体对比

四、PPO vs TRPO vs A2C对比

4.1 算法对比表

4.2 为什么PPO成为主流？

4.3 PPO2与向量化环境

五、实战调参经验

5.1 超参数推荐值

5.2 调试技巧

5.3 常见问题与解决

六、数学形式化总结

PPO剪裁目标函数

自适应KL惩罚目标

PPO完整目标

七、相关文档

参考文献

关系图谱

目录

反向链接